Browsers like Chrome, Chromium and Mozilla Firefox have moved to running tabs in separate threads and processes, to increase performance and responsiveness and to reduce the effects of crashes in one tab.

Occasionally, this exhausts the default limit on the amount of processes and threads that a user can have running.

Determine the maximum number of processes and threads in a user session:

$ ulimit -u
1200

The SUSE defaults are configured in /etc/security/limits.conf:

# harden against fork-bombs
* hard nproc 1700
* soft nproc 1200
root hard nproc 3000
root soft nproc 1850

In the above, * the catch-all for all users.

To raise the limit for a particular user, you can either edit /etc/security/limits.conf or create a new file /etc/security/limits.d/nproc.conf. Here is an example for /etc/security/limits.d/nproc.conf raising the limit for the user jdoe to 8k/16k threads and processes:

jdoe soft nproc 8192
jdoe hard nproc 16384

If you want to do that for a whole group, use the @ prefix:

@powerusers soft nproc 8192
@powerusers hard nproc 16384

In either case, this change is effective only for the next shell or login session.

I just realized that I only tweeted about this a couple of months ago, but never blogged about it. Shame on me!

I wrote an article, Legacy Systems as Old Cities, for The Recompiler magazine. Is GNOME, now at 20 years old, legacy software? Is it different from mainframe software because "everyone" can change it? Does long-lived software have the same patterns of change as cities and physical artifacts? Can we learn from the building trades and urbanism for maintaining software in the long term? Could we turn legacy software into a good legacy?

You can read the article here.

Also, let me take this opportunity to recommend The Recompiler magazine. It is the most enjoyable technical publication I read. Their podcast is also excellent!

Update 2017/06/10 - Spanish version of the article, Los Sistemas Heredados como Ciudades Viejas

After updating my distro a few months ago, I somehow lost my tweaks to the Alt-Tab behavior in gnome-shell.

The default is to have Alt-Tab switch you between applications in the current workspace. One can use Alt-backtick (or whatever key you have above Tab) to switch between windows in the current application.

I prefer a Windows-like setup, where Alt-Tab switches between windows in the current workspace, regardless of the application to which they belong.

Many moons ago there was a gnome-shell extension to change this behavior, but these days (GNOME 3.24) it can be done without extensions. It is a bit convoluted.

With the GUI

If you are using X instead of Wayland, this works:

1. Unset the Switch applications command. To do this, run gnome-control-center, go to Keyboard, and find the Switch applications command. Click on it, and hit Backspace in the dialog that prompts you for the keyboard shortcut. Click on the Set button.

2. Set the Switch windows command. While still in the Keyboard settings, find the Switch windows command. Click on it, and hit Alt-Tab. Click Set.

That should be all you need, unless you are in Wayland. In that case, you need to do it on the command line.

With the command line, or in Wayland

The kind people on #gnome-hackers tell me that as of GNOME 3.24, changing Alt-Tab doesn't work on Wayland as in (2) above, because the compositor captures the Alt-Tab key when you type it inside the dialog that prompts you for a keyboard shortcut. In that case, you have to change the configuration keys directly instead of using the GUI:

gsettings set org.gnome.desktop.wm.keybindings switch-applications "[]"
gsettings set org.gnome.desktop.wm.keybindings switch-applications-backward "[]"
gsettings set org.gnome.desktop.wm.keybindings switch-windows "['<Alt>Tab', '<Super>Tab']"
gsettings set org.gnome.desktop.wm.keybindings switch-windows-backward  "['<Alt><Shift>Tab', '<Super><Shift>Tab']"

Of course the above also works in X, too.

Changing windows across all workspaces

If you'd like to switch between windows in all workspaces, rather than in the current workspace, find the org.gnome.shell.window-switcher current-workspace-only GSettings key and change it. You can do this in dconf-editor, or on the command line with

gsettings set org.gnome.shell.window-switcher current-workspace-only true

In this post I'll show you the code path that Rust takes inside its standard library when you open a file. I wanted to learn how Rust handles system calls and errno, and all the little subtleties of the POSIX API. This is what I learned!

The C side of things

When you open a file, or create a socket, or do anything else that returns an object that can be accessed like a file, you get a file descriptor in the form of an int.

/* All of these return a int with a file descriptor, or
 * -1 in case of error.
 */
int open(const char *pathname, int flags, ...);
int socket(int domain, int type, int protocol);

You get a nonnegative integer in case of success, or -1 in case of an error. If there's an error, you look at errno, which gives you an integer error code.

int fd;

retry_open:
fd = open ("/foo/bar/baz.txt", 0);
if (fd == -1) {
    if (errno == ENOENT) {
        /* File doesn't exist */
    } else if (errno == ...) [
        ...
    } else if (errno == EINTR) {
        goto retry_open; /* interrupted system call; let's retry */
    }
}

Many system calls can return EINTR, which means "interrupted system call", which means that something interrupted the kernel while it was doing your system call and it returned control to userspace, with the syscall unfinished. For example, your process may have received a Unix signal (e.g. you send it SIGSTOP by pressing Ctrl-Z on a terminal, or you resized the terminal and your process got a SIGWINCH). Most of the time EINTR means simply that you must retry the operation: if you Control-Z a program to suspend it, and then fg to continue it again; and if the program was in the middle of open()ing a file, you would expect it to continue at that exact point and to actually open the file. Software that doesn't check for EINTR can fail in very subtle ways!

Once you have an open file descriptor, you can read from it:

ssize_t
read_five_bytes (int fd, void *buf)
{
    ssize_t result;

    retry:
    result = read (fd, buf, 5);
    if (result == -1) {
        if (errno == EINTR) {
            goto retry;
        } else {
            return -1; /* the caller should cherk errno */
        }
    } else {
        return result; /* success */
    }
}

... and one has to remember that if read() returns 0, it means we were at the end-of-file; if it returns less than the number of bytes requested it means we were close to the end of file; if this is a nonblocking socket and it returns EWOULDBLOCK or EAGAIN then one must decide to retry the operation or actually wait and try again later.

There is a lot of buggy software written in C that tries to use the POSIX API directly, and gets these subtleties wrong. Most programs written in high-level languages use the I/O facilities provided by their language, which hopefully make things easier.

I/O in Rust

Rust makes error handling convenient and safe. If you decide to ignore an error, the code looks like it is ignoring the error (e.g. you can grep for unwrap() and find lazy code). The code actually looks better if it doesn't ignore the error and properly propagates it upstream (e.g. you can use the ? shortcut to propagate errors to the calling function).

I keep recommending this article on error models to people; it discusses POSIX-like error codes vs. exceptions vs. more modern approaches like Haskell's and Rust's - definitely worth studying over a few of days (also, see Miguel's valiant effort to move C# I/O away from exceptions for I/O errors).

So, what happens when one opens a file in Rust, from the toplevel API down to the system calls? Let's go down the rabbit hole.

You can open a file like this:

use std::fs::File;

fn main () {
    let f = File::open ("foo.txt");
    ...
}

This does not give you a raw file descriptor; it gives you an io::Result<fs::File, io::Error>, which you must pick apart to see if you actually got back a File that you can operate on, or an error.

Let's look at the implementation of File::open() and File::create().

impl File {
    pub fn open<P: AsRef<Path>>(path: P) -> io::Result<File> {
        OpenOptions::new().read(true).open(path.as_ref())
    }

    pub fn create<P: AsRef<Path>>(path: P) -> io::Result<File> {
        OpenOptions::new().write(true).create(true).truncate(true).open(path.as_ref())
    }
    ...
}

Here, OpenOptions is an auxiliary struct that implements a "builder" pattern. Instead of passing bitflags for the various O_CREATE/O_APPEND/etc. flags from the open(2) system call, one builds a struct with the desired options, and finally calls .open() on it.

So, let's look at the implementation of OpenOptions.open():

    pub fn open<P: AsRef<Path>>(&self, path: P) -> io::Result<File> {
        self._open(path.as_ref())
    }

    fn _open(&self, path: &Path) -> io::Result<File> {
        let inner = fs_imp::File::open(path, &self.0)?;
        Ok(File { inner: inner })
    }

See that fs_imp::File::open()? That's what we want: it's the platform-specific wrapper for opening files. Let's look at its implementation for Unix:

    pub fn open(path: &Path, opts: &OpenOptions) -> io::Result<File> {
        let path = cstr(path)?;
        File::open_c(&path, opts)
    }

The first line, let path = cstr(path)? tries to convert a Path into a nul-terminated C string. The second line calls the following:

    pub fn open_c(path: &CStr, opts: &OpenOptions) -> io::Result<File> {
        let flags = libc::O_CLOEXEC |
                    opts.get_access_mode()? |
                    opts.get_creation_mode()? |
                    (opts.custom_flags as c_int & !libc::O_ACCMODE);
        let fd = cvt_r(|| unsafe {
            open64(path.as_ptr(), flags, opts.mode as c_int)
        })?;
        let fd = FileDesc::new(fd);

        ...

        Ok(File(fd))
    }

Here, let flags = ... converts the OpenOptions we had in the beginning to an int with bit flags.

Then, it does let fd = cvt_r (LAMBDA), and that lambda function calls the actual open64() from libc (a Rust wrapper for the system's libc): it returns a file descriptor, or -1 on error. Why is this done in a lambda? Let's look at cvt_r():

pub fn cvt_r<T, F>(mut f: F) -> io::Result<T>
    where T: IsMinusOne,
          F: FnMut() -> T
{
    loop {
        match cvt(f()) {
            Err(ref e) if e.kind() == ErrorKind::Interrupted => {}
            other => return other,
        }
    }
}

Okay! Here f is the lambda that calls open64(); cvt_r() calls it in a loop and translates the POSIX-like result into something friendly to Rust. This loop is where it handles EINTR, which gets translated into ErrorKind::Interrupted. I suppose cvt_r() stands for convert_retry()? Let's look at the implementation of cvt(), which fetches the error code:

pub fn cvt<T: IsMinusOne>(t: T) -> io::Result<T> {
    if t.is_minus_one() {
        Err(io::Error::last_os_error())
    } else {
        Ok(t)
    }
}

(The IsMinusOne shenanigans are just a Rust-ism to help convert multiple integer types without a lot of as casts.)

The above means, if the POSIX-like result was -1, return an Err() from the last error returned by the operating system. That should surely be errno internally, correct? Let's look at the implementation for io::Error::last_os_error():

    pub fn last_os_error() -> Error {
        Error::from_raw_os_error(sys::os::errno() as i32)
    }

We don't need to look at Error::from_raw_os_error(); it's just a conversion function from an errno value into a Rust enum value. However, let's look at sys::os::errno():

pub fn errno() -> i32 {
    unsafe {
        (*errno_location()) as i32
    }
}

Here, errno_location() is an extern function defined in GNU libc (or whatever C library your Unix uses). It returns a pointer to the actual int which is the errno thread-local variable. Since non-C code can't use libc's global variables directly, there needs to be a way to get their addresses via function calls - that's what errno_location() is for.

And on Windows?

Remember the internal File.open()? This is what it looks like on Windows:

    pub fn open(path: &Path, opts: &OpenOptions) -> io::Result<File> {
        let path = to_u16s(path)?;
        let handle = unsafe {
            c::CreateFileW(path.as_ptr(),
                           opts.get_access_mode()?,
                           opts.share_mode,
                           opts.security_attributes as *mut _,
                           opts.get_creation_mode()?,
                           opts.get_flags_and_attributes(),
                           ptr::null_mut())
        };
        if handle == c::INVALID_HANDLE_VALUE {
            Err(Error::last_os_error())
        } else {
            Ok(File { handle: Handle::new(handle) })
        }
    }

CreateFileW() is the Windows API function to open files. The conversion of error codes inside Error::last_os_error() happens analogously - it calls GetLastError() from the Windows API and converts it.

Can we not call C libraries?

The Rust/Unix code above depends on the system's libc for open() and errno, which are entirely C constructs. Libc is what actually does the system calls. There are efforts to make the Rust standard library not use libc and use syscalls directly.

As an example, you can look at the Rust standard library for Redox. Redox is a new operating system kernel entirely written in Rust. Fun times!

Update: If you want to see what a C-less libstd would look like, take a look at steed, an effort to reimplement Rust's libstd without C dependencies.

Conclusion

Rust is very meticulous about error handling, but it succeeds in making it pleasant to read. I/O functions give you back an io::Result<>, which you piece apart to see if it succeeded or got an error.

Internally, and for each platform it supports, the Rust standard library translates errno from libc into an io::ErrorKind Rust enum. The standard library also automatically handles Unix-isms like retrying operations on EINTR.

I've been enjoying reading the Rust standard library code: it has taught me many Rust-isms, and it's nice to see how the hairy/historical libc constructs are translated into clean Rust idioms. I hope this little trip down the rabbit hole for the open(2) system call lets you look in other interesting places, too.

It’s an idiom that quickly becomes rote to Go programmers: whenever you conjure up a value that implements the io.Closer interface, after checking for errors you immediately defer its Close() method. You see this most often when making HTTP requests:

resp, err := http.Get("https://joeshaw.org")
if err != nil {
    return err
}
defer resp.Body.Close()

or opening files:

f, err := os.Open("/home/joeshaw/notes.txt")
if err != nil {
    return err
}
defer f.Close()

But this idiom is actually harmful for writable files because deferring a function call ignores its return value, and the Close() method can return errors. For writable files, Go programmers should avoid the defer idiom or very infrequent, maddening bugs will occur.

Why would you get an error from Close() but not an earlier Write() call? To answer that we need to take a brief, high-level detour into the area of computer architecture.

Generally speaking, as you move outside and away from your CPU, actions get orders of magnitude slower. Writing to a CPU register is very fast. Accessing system RAM is quite slow in comparison. Doing I/O on disks or networks is an eternity.

If every Write() call committed the data to the disk synchronously, the performance of our systems would be unusably slow. While synchronous writes are very important for certain types of software (like databases), most of the time it’s overkill.

The pathological case is writing to a file one byte at a time. Hard drives – brutish, mechanical devices – need to physically move a magnetic head to the position on the platter and possibly wait for a full platter revolution before the data could be persisted. SSDs, which store data in blocks and have a finite number of write cycles for each block, would quickly burn out as blocks are repeatedly written and overwritten.

Fortunately this doesn’t happen because multiple layers within hardware and software implement caching and write buffering. When you call Write(), your data is not immediately being written to media. The operating system, storage controllers and the media itself are all buffering the data in order to batch smaller writes together, organizing the data optimally for storage on the medium, and deciding when best to commit it. This turns our writes from slow, blocking synchronous operations to quick, asynchronous operations that don’t directly touch the much slower I/O device. Writing a byte at a time is never the most efficient thing to do, but at least we are not wearing out our hardware if we do it.

Of course, the bytes do have to be committed to disk at some point. The operating system knows that when we close a file, we are finished with it and no subsequent write operations are going to happen. It also knows that closing the file is its last chance to tell us something went wrong.

On POSIX systems like Linux and macOS, closing a file is handled by the close system call. The BSD man page for close(2) talks about the errors it can return:

ERRORS
     The close() system call will fail if:

     [EBADF]            fildes is not a valid, active file descriptor.

     [EINTR]            Its execution was interrupted by a signal.

     [EIO]              A previously-uncommitted write(2) encountered an input/output
                        error.

EIO is exactly the error we are worried about. It means that we’ve lost data trying to save it to disk, and our Go programs should absolutely not return a nil error in that case.

The simplest way to solve this is simply not to use defer when writing files:

func helloNotes() error {
    f, err := os.Create("/home/joeshaw/notes.txt")
    if err != nil {
        return err
    }

    if err = io.WriteString(f, "hello world"); err != nil {
        f.Close()
        return err
    }

    return f.Close()
}

This does mean additional bookkeeping of the file in the case of errors: we must explicitly close it in the case where io.WriteString() fails (and ignore its error, because the write error takes precedence). But it’s clear, straightforward, and properly checks the error from the f.Close() call.

There is a way to handle this case with defer by using named return values and a closure:

func helloNotes() (err error) {
    var f *os.File
    f, err = os.Create("/home/joeshaw/notes.txt")
    if err != nil {
        return
    }

    defer func() {
        cerr := f.Close()
        if err == nil {
            err = cerr
        }
    }()

    err = io.WriteString(f, "hello world")
    return
}

The main benefit of this pattern is that it’s not possible to forget to close the file because the deferred closure always executes. In longer functions with more if err != nil conditional branches, this pattern can also result in fewer lines of code and less repetition.

Still, I find this pattern to be a little too magical. I dislike using named return values, and modifying the return value after the core function finishes is not intuitively clear even to experienced Go programmers.

I am willing to accept the tradeoff of more readable and easily understandable code for needing to obsessively review code to ensure that the file is closed in all cases, and that’s the approach I recommend in code reviews I give to others.

Update

On Twitter, Ben Johnson suggested that Close() may be safe to run multiple times on files, like so:

func doSomething() error {
    f, err := os.Create("foo")
    if err != nil {
        return err
    }
    defer f.Close()

    if _, err := f.Write([]byte("bar"); err != nil {
        return err
    }

    if err := f.Close(); err != nil {
        return err
    }
    return nil
}

(gist)

The Go docs on io.Closer explicitly say that at an interface level behavior after the first call is unspecificed, but specific implementations may document their own behavior.

The docs for *os.File unfortunately aren’t clear on its behavior, saying only, “Close closes the File, rendering it unusable for I/O. It returns an error, if any.” The implemenation as of 1.8, however, shows:

func (f *File) Close() error {
    if f == nil {
        return ErrInvalid
    }
    return f.file.close()
}

func (file *file) close() error {
    if file == nil || file.fd == badFd {
        return syscall.EINVAL
    }
    var err error
    if e := syscall.Close(file.fd); e != nil {
        err = &PathError{"close", file.name, e}
    }
    file.fd = -1 // so it can't be closed again

    // no need for a finalizer anymore
    runtime.SetFinalizer(file, nil)
    return err
}

For clarity, badFd is defined as -1, so subsequent attempts to close an *os.File will do nothing and return syscall.EINVAL. But since we are ignoring the error from the defer, this doesn’t matter. It’s not idempotent, exactly, but as Ben put later in the Twitter thread, it “won’t blow shit up if you call it twice.”

The implementation is a good, common-sense one and it seems unlikely to change in the future and cause problems. But the lack of documentation about this outcome makes me a little nervous. Maybe a doc update to codify this behavior would be a good task for Go 1.10.

Update 2

Closing the file is the last chance the OS has to tell us about problems, but the buffers are not necessarily going to be flushed when you close the file. It’s entirely possible that flushing the write buffer to disk will happen after you close the file, and a failure there cannot be caught. If this happens, it usually means you have something seriously wrong, like a failing disk.

However, you can force the write to disk with the Sync() method on *os.File, which calls the fsync system call. You should check for errors from that call, but then I think it’s safe to ignore an error from Close(). Calling fsync has serious implications on performance: it’s flushing write buffers out to slow disks. But if you really, really want the data on disk, the best pattern to follow is probably:

func helloNotes() error {
    f, err := os.Create("/home/joeshaw/notes.txt")
    if err != nil {
        return err
    }
    defer f.Close()

    if err = io.WriteString(f, "hello world"); err != nil {
        return err
    }

    return f.Sync()
}